It takes significant skill to generate huge apps.

I'm very curious to hear folks' thoughts on this. My intuition is that there are at least two key factors that make tasks difficult for AI tools.

(1.) The overall level of complexity. For example, I think (today) you would need to do an extreme amount of hand-holding to get an AI to write a browser, an OS, or anything with similarly very high level of complexity.

(2.) The representation in the AI's training set. Anything that is common in the training set (e.g., to-do app in JavaScript) will be trivial for an AI to one-shot. Anything that is rare, or absent in the training set is going to be much, much harder for the AI. If I try to ask an AI tool to develop a automated, ultra-low-latency high-frequency trading system, it's probably going to struggle because those kinds of applications aren't in the open-source domain. The same is true for completely novel algorithms. So maybe things that are essentially new science/engineering.

I'm very curious to hear others' thoughts on this, as I've been wondering about this, too.

From what I’ve seen, AI tools don’t do well with large codebases with lots of technical debt and complex interdependencies. If you see some project that requires tons of manual testing to keep running and only a few crusty veterans know how to make changes without causing more problems, this is the kind of project that could be fairly immune from AI takeover. Bonus points if there are poor or limited unit tests. Projects like these exist in the bowels of many large companies chugging along helping keep the business running. They may not be sexy, but if you really want to avoid AI tools, it could be an option

Security / Finance. People aren't going to trust slop to make critical decisions for a long time.

Any niche areas - once you venture outside of topics/tasks well represented on the Internet and GitHub, even frontier LLMs still quickly take a nosedive.

Not totally safe, but embedded systems is a nice pivot. The programming difficulty is higher (imo), but not many generative physical hardware systems.

Medical and defense, for regulatory and security reasons respectively.